Extreme Model Compression for On-device Natural Language Understanding